An answer can name a brand without showing a citation beside the statement. Record the mention and its context. Do not invent a source URL or assume that a recent visit in your server logs explains where the answer came from.
The investigation should separate three questions: what the answer says, whether the statement is accurate, and what source or mechanism produced it. You may be able to answer the first two while the third remains unknown.
This is an original diagnostic method. It does not claim access to an engine's hidden retrieval process or imply that every unlinked answer is generated in the same way.
Build a short observation record
Preserve the exact prompt, relevant conversation context, answer, observation date, interface, and visible links. Record whether the brand appears in the answer body, a source panel, or another surface.
| Question | Permitted conclusion from the record |
|---|---|
| Was the product named? | Yes or no, after identity review |
| Was it recommended? | Only according to the stated coding rule and context |
| Was a citation visible? | Record the actual destination or no visible citation |
| Where did the engine obtain the fact? | Unknown unless the available evidence establishes it |
“No visible citation” is more precise than “no source exists.” The interface may not expose everything that influenced the answer. At the same time, a hidden-source possibility is not evidence for any particular source.
Check the claim against current product evidence
If the answer says your product supports a workflow, compare that statement with the current approved product record. Record supported, partial, contradicted, or unresolved status using the same evidence discipline applied to linked claims.
In a hypothetical example, an answer names a fictional product and says every plan includes export. The product's current plan page limits export to one plan. The accuracy problem is clear even though the answer does not reveal its source.
Correct your own public information if it is inconsistent. If your pages already state the limit clearly, retain the evidence and use an appropriate feedback channel when available. Do not promise that a correction request will change future answers.
Avoid retrospective source storytelling
Finding a page that contains similar wording does not prove that the engine used that page. A crawler request shortly before the observation also does not establish that the fetched resource caused the mention.
Record those observations as possible leads with their limitations. Keep chronology, textual similarity, and verified provenance distinct. A neat story can be more persuasive than the evidence warrants.
If the product exposes a source-inspection feature, use its current documented behavior and preserve the actual result. Do not generalize that feature's evidence to other interfaces or past observations where it was not used.
Choose a useful next action
An accurate unlinked mention may simply remain a mention observation. An inaccurate one may justify a clearer product explanation, a source consistency review, or a documented correction attempt. A missing link does not automatically imply a technical defect on your site.
Use the claim-support method for accuracy and the capture manifest for the evidence record. Report what is known and leave the hidden provenance unresolved when the available tools cannot establish it.